Objective
DiaMant aims to help prevent hypoglycemic events in people with type 1 diabetes by predicting risk up to 120 minutes before onset. Through an app-based interface, the system provides timely alerts and intervention recommendations tailored to the predicted time horizon. A key objective of the project is to evaluate the role of personalization in improving hypoglycemia prediction and evaluate whether population-based models can generalize across subjects and age groups.
Description
Type 1 diabetes is an autoimmune disease that requires insulin therapy to regulate blood glucose levels. However, insulin treatment can lower glucose levels below 70 mg/dL, leading to hypoglycemia. If not addressed in time, hypoglycemia may cause dizziness, loss of consciousness, coma, and, in severe cases, death. Early detection can support preventive actions such as eating a snack or resting.
DiaMant addresses this challenge by classifying short-term hypoglycemia onset. The project aims to improve predictive performance and to overcome a key limitation in current research: the lack of sufficiently large, well-curated datasets that support robust analysis across broader patient populations rather than isolated study cohorts.
Current Research
To support model development, we curated DiaData by integrating 15 publicly available datasets. DiaData includes data from 2510 people with type 1 diabetes and combines continuous glucose monitoring data, heart rate values, demographic information, laboratory values, and personal health features, including age, sex, BMI, height, weight, and race [1]. The dataset was further improved using advanced data imputation and quality enhancement modules, applied individually to specific time ranges of gaps [2]. The system classifies hypoglycemia risk across clinically relevant prediction horizons: at onset and 5-15, 20-45, and 50-120 minutes before onset. These time windows support actionable interventions, particularly during the 5-15 minutes before onset, when self-treatment may still be possible [3].
To improve robustness, we developed multiple model variants, including versions with and without heart rate data. The classification models were incorporated into a real-time application framework that supports data collection, model execution, and intervention recommendations based on the predicted time to hypoglycemia. For example, if risk is detected within 0-25 minutes, the app recommends sitting down and consuming a snack [4].
In addition, we assessed whether adding personal features or vital signs to glucose time-series data improves model performance. We also trained specialized expert models for different age groups: children aged 2-13 years, teenagers aged 14-20 years, adults aged 21-44 years, and older adults aged 45+ years. These models were compared with a population-based model to evaluate generalizability across age groups, showing that children benefit from specialized expert models [3].
Future Work
Using DiaData, we will develop and compare a wide range of machine learning and deep learning models. To further improve prediction performance and interpretability, we will also investigate ensemble learning and conduct SHAP-based feature analysis.
References
[1] Cinar B, Maleshkova M. Benchmarking hypoglycemia classification using quality-enhanced DiaData. IEEE J Biomed Health Inform. 2025;29(12):8831-8838. doi:10.1109/JBHI.2025.3620603
[2] Gupta V, Grensing F, Cinar B, Van Den Boom L, Maleshkova M. Fill in the gaps – applying polynomial-based imputation techniques for heart rate data. In: 2026 IEEE Conference on Artificial Intelligence (CAI). IEEE; 2026:1174-1181. doi:10.1109/cai68641.2026.11536361
[3] Cinar B, Maleshkova M. Evaluating generalizability of population-based and age-segmented models for hypoglycemia classification. In: 2026 IEEE Conference on Artificial Intelligence (CAI). IEEE; 2026:1348-1355. doi:10.1109/cai68641.2026.11536643
[4] Grensing F, Cinar B, Maleshkova M. Early warning of hypoglycemia via sensor-agnostic machine learning: a clinical app design for type 1 diabetes. In: International Conferences on Applied Computing 2025 and WWW/Internet 2025: Proceedings. IADIS Press; 2025. Presented at: 22nd International Conference on Applied Computing 2025 and 24th International Conference on WWW/Internet 2025 (AC ICWI 2025); November 1-3, 2025; Porto, Portugal. support robust analysis across broader patient populations rather than isolated study cohorts.
Letzte Änderung: 6. August 2026